Agent skill

A reusable set of saved instructions that an AI agent loads and follows when it recognizes a matching task, like a recipe card an experienced cook pulls from a drawer instead of reinventing the dish each time.

An agent skill is a reusable set of saved instructions that an AI agent loads and follows whenever it recognizes a matching task. You describe the workflow once, including what the agent should do, what inputs to expect, and what the finished output looks like. From then on, the agent pulls those instructions into any conversation where the same kind of work comes up. In Codex, a skill lives in a SKILL.md file inside a folder. In Claude, skills are packaged as instruction bundles attached to a project. The principle is the same across platforms: capture a proven workflow once so the agent can repeat it reliably.

Builder example

Without skills, every conversation starts from zero. You re-explain the same formatting rules, the same review checklist, the same output structure. Skills eliminate that repetition and make the agent's behavior consistent across conversations. They also make your workflows portable: a skill can be shared with teammates, version-controlled in a repository, or moved between projects.

Common confusion: A skill is often confused with a single prompt or a system instruction. A prompt is what you type in one conversation. A system instruction sets general behavior for all conversations. A skill sits between them: it activates only when the task matches, and it carries detailed step-by-step instructions specific to one workflow.